Cheminformatics

TxBench-PP: Analyzing AI Agent Performance on Small-Molecule Preclinical Pharmacology

Hannah Le, Ramesh Ramasamy, Alex Urrutia, Mahsa Yazdani, Tim Proctor, Kenny Workman

Featured June 20, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

A new test, TxBench-PP, helps see if AI can make smart drug decisions from real lab data, showing that even the best AI still struggles to act like a real scientist.

In depth

The paper introduces TxBench-PP, a novel benchmark designed to rigorously evaluate AI agents on small-molecule preclinical pharmacology decisions. It provides agents with realistic experimental data and workflow context, requiring them to recover accurate conclusions rather than relying on memorized literature. The benchmark employs deterministic grading of structured answers, revealing that current AI systems struggle with scientific judgment in this complex domain.

Key Takeaways

1
TxBench-PP is a new, verifiable benchmark for AI agents in small-molecule preclinical pharmacology, focusing on data-driven decision-making.
2
The benchmark comprises 100 evaluations across diverse program stages, assay types, and task structures, designed to prevent memorization.
3
Current AI agents, even top configurations, perform poorly, with the best achieving only 59.3% accuracy, highlighting significant gaps in scientific judgment.

Conceptual Flow

HIGH LEVEL

Methodology: How AI Agents are Tested

The paper created a special test for AI by giving it fake lab data and asking it to make drug decisions, just like a real scientist would.

Real Lab Data

AI Agent Thinks

Drug Decision

Results: AI Agents Struggle

The test showed that even the smartest AI agents often make mistakes, proving they are not yet good enough to make important drug decisions on their own.

AI Agents

Struggle with Decisions

Many Mistakes

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.